blind image quality assessment with local ...aircconline.com/ijci/v5n4/5416ijci22.pdfpg scholar,...

International Journal on Cybernetics & Informatics (IJCI) Vol. 5, No. 4, August 2016

DOI: 10.5121/ijci.2016.5422 191

BLIND IMAGE QUALITY ASSESSMENT WITH LOCAL

CONTRAST FEATURES

Ganta Kasi Vaibhav, PG Scholar, Department of Electronics and Communication Engineering,

University College of Engineering Vizianagaram,JNTUK.

Ch.Srinivasa Rao, Professor, Department of Electronics and Communication Engineering,

University College of Engineering Vizianagaram,JNTUK.

ABSTRACT

The aim of this research is to create a tool to evaluate distortion in images without the information about

original image. Work is to extract the statistical information of the edges and boundaries in the image and

to study the correlation between the extracted features. Change in the structural information like shape and

amount of edges of the image derives quality prediction of the image. Local contrast features are effectively

detected from the responses of Gradient Magnitude (G) and Laplacian of Gaussian (L) operations. Using

the joint adaptive normalisation, G and L are normalised. Normalised values are quantized into M and N

levels respectively. For these quantised M levels of G and N levels of L, Probability (P) and conditional

probability(C) are calculated. Four sets of values namely marginal distributions of gradient magnitude Pg,

marginal distributions of Laplacian of Gaussian Pl, conditional probability of gradient magnitude Cg and

probability of Laplacian of Gaussian Cl are formed. These four segments or models are Pg, Pl, Cg and Cl.

The assumption is that the dependencies between features of gradient magnitude and Laplacian of

Gaussian can formulate the level of distortion in the image. To find out them, Spearman and Pearson

correlations between Pg, Pl and Cg, Cl are calculated. Four different correlation values of each image are

the area of interest. Results are also compared with classical tool Structural Similarity Index Measure

(SSIM)

KEYWORDS

Gradient Magnitude, Laplacian of Gaussian, Joint Adaptive Normalisation, Normalised Bivariate

Histograms, Spearman rank Correlation, Pearson Correlation Coefficient.

1. INTRODUCTION

Image quality assessment evaluates the quality of the distorted image. Factors which determine

image quality are, noise, dynamic range tone reproduction, colour accuracy, distortion, contrast,

exposure accuracy, lateral chromatic aberration, sharpness, colour moiré, vignette, artefacts.

Distortion is defined as abnormality, irregularity or variation caused in an image. This is

noticeable in low cost cameras. Distortions are caused during Acquisition, Compression,

Transmission and Storage. Changes in image or quality of image are observed either by the

human subjects called as subjective measure or calculated by mathematical operations called as


192

objective measures. Image quality assessment can also be categorised as With Reference models

and Without Reference models. First type of models finds out quality of the image by comparing

with its original image. Second type of models also called as Blind image quality assessment

finds out the quality of distorted image without comparing with its original image. Local contrast features describe the structure of the image. The changes in the structure of the

image like shape and amount of edges are detected easily. Two general local contrast features are

Gradient magnitude and Laplacian of Gaussian. Joint adaptive normalisation (JAN) normalises G

and L channels jointly. The benefit of JAN is to make the horizontal, vertical and diagonal

features correlative in the image. It reduces the redundancies in image. Normalisation stabilizes

the profiles of these features. The G and L distributions of images are very different from natural

images. Quantising into levels and thereby giving joint probability function, statistics are derived.

Marginal distributions and conditional probability dependency measures are recorded into four

different correlations are drawn. Over years there had been several performance measures for image quality assessment.

DMOS/MOS difference mean opinion scores is a subjective measure which evaluates over human

judgements [4]. Various databases established by IQA community are LIVE, CSIQ, TID2008.

CSIQ, TID2008 and LIVE has 4 common types of distortions they are JP2K, JPEG, WN and

Gaussian blur. These are used to identify the characteristics of the various distortions. To find out

performance of a method, a machine which calculates the correlations of the subjective scores of

the human judgements constructed. The correlations used are Spearman rank order correlation

coefficient (SRC) and Pearson correlation coefficient (PCC). Proposed Blind image quality

assessment model is compared with Structural Similarity Index Measure (SSIM). Experimental results define that there is equivalent information in one of the four sets of statistics

we derive. However joint statistics in of Pearson model give better results. Existing models

involve in large procedures to find out the disturbances in the frequencies over different

distortions. Few models even changes the features of the image. G and L operations are close to

the results of the human visual system. G and L are independent over the distortions. The

marginal and independency distributions can determine the quality of the image. Joint adaptive

normalisation procedure normalises the G and L features. Proposed model uses independency

distributions to measure joint statistics. Which leads to highly competitive results in terms of

Quality prediction, Generalisation ability, and effectiveness. Existing models have computational

complexity. To reduce regression methods, and get finer quality measure with no training or

regression methods we derive a new method with probability statistics. This paper is organised in four sections. Section II gives a brief study of all the existing models of

without reference image quality assessment. Section III presents the features of the proposed

model in detail and exclusive experimental results in each stage of the process. Section IV

concludes the paper.

2. RELATED WORK

2.1.Literature Survey

There are several image quality assessment models. Mean Square Error (MSE) is the primitive

measure [2]. When original image is the known reference image, a written explanation of the

measure exits in the literature. Mean Square Error (MSE) found to be very important measure to


193

compare two signals. It provides a similarity index score that gives the degree of similarity.

Similarity index map is amount of distortion between two signals. It is simple, parameter free and

inexpensive. It is employed widely for optimizing and assessing signal processing applications.

Yet it did not measure signal fidelity to certain required extinct. When altered by two different

distortions at a same level, MSE only gave the value of distortion. MSE is very converse to the

human perception. MSE led to development of Minimum Mean Square Error, Peak Signal to

Noise Ratio. Based on the luminance, contrast and structure of an image, the Structural Similarity Index

Measure (SSIM) [3] is developed. It is an objective quality measure of image to quantify the

visibility of errors. It is a similarity measure for comparing any two signals. Like MSE, it is a full

reference model which depends on the structural similarity. It has SSIM index map which shows

the comparison better than MSE. It is a complex measure for an image with large content.

Though SSIM index could give better results compared to the traditional Mean square error, it

lacked in defining what the type of distortion is present in the image. This novel work gave scope

for the study on structure statistics of the image. Furthermore, it is used to simplify algorithms of

image processing system. For image quality measure, SSIM proved its efficiency than mean

square error. SSIM is computationally expensive than MSE. Difference Mean Opinion Scoring (DMOS) is a performance evaluation study of existing

methods of IQA along with subjective scores collected over a period of time with numerous

human subjects [4]. There is no replacement to the Human Visual system (HVS) [5]. DMOS

when compared objective measures like PSNR, MSE and SSIM, proved that Human Visual

System had better evaluation. Subjective group verified and gave responses over LIVE database

which consists of 779 distorted images from 29 original images with five distortions [14]. It gave

importance to visual difference. Used of Spearman Rank Correlation and Root Mean Square

Error for perfection. It has 95% confidence criterion of finding out whether distorted or not. This

study gave a valuable resource of scores of distortion. It led to study of natural scene statistics.

DMOS is used as benchmark for checking of constructed models. The reference image is not always available. Hence, there is need of without reference image

quality assessment measure also called as Blind Image Quality Assessment. It assesses the quality

of an image completely blind i.e., without any knowledge of source distortion [5]. Distorted

Image Statistics (DIS) which is used to classify images into distortion categories gave the ease to

decide type of the distortion [5]. In this literature, wavelet transform is performed on the image

indices. Shape parameter is defined using Gaussian distribution. Given a training set and testing

setoff distorted images, a classifier Support Vector Machine is used to classify image into five

distortions. This is the first without reference method. It could differentiate the type of distortion.

Results of this model correlates with reference models. Its drawback is its computational

complexity. This methodology can be replaced with any module which performs better i.e. either

by increasing the number of distortions or by increasing training set for better results. This model

can be used for video processing by adding measure of relevant perceptual features. With the probability of usage of the indices as features, a new approach BLIINDS is proposed

[5]. It is a model with the evolution of features derived from the Discrete Cosine Transform

domain statistics. While the previous no reference is distorted specific approaches, this approach

could explain the type of the distortion. It used Support Vector Machine (SVM) which correlates

well with human visual perception. It is computationally convenient as it is based on a DCT-

framework entirely, and beats the performance of Peak signal to noise ratio. The probabilistic

prediction model was trained on a small sample of the data, and only required the computation of


194

the mean and the covariance of the training data. It computes blockiness measure. It estimates the

particular type of distortion. Time taken for computation is high. Final evolution is the reduction of complexity by combination with Natural scene statistics and

addition of distorted image statistics. Work is to extract the statistical information of the edges

and boundaries in the image and to study the correlation between the extracted features. Change

in the structural information like shape and amount of edges of the image derives quality

prediction of the image. Local contrast features are effectively detected from the responses of

Gradient Magnitude (G) and Laplacian of Gaussian (L) operations. Adaptive procedures are used

to normalize the values of G and L. Normalized values are quantized into certain levels

respectively. Conditional probability and Marginal distribution of G and L are calculated which

are stored into three segments. They proposed three models. These three segments or models are

M1, M2, M3 which have only conditional probability values, only marginal distribution values

and both conditional probability and marginal distribution values. Loading values in the support

vector regression; over the set of images collected from LIVE database a probable score is

determined for each distortion. There is complexity in training these values into the support vector machine. This paper is a

thorough study of the conditional probability and marginal probability values of the gradient

magnitude and Laplacian of Gaussian. Marginal distributions and conditional probabilities and

their dependencies which lead to highly competitive performance are employed in this work. This

avoided the training and learning of the features. Therefore, the complexity is reduced in the

modelling. This procedure is a direct extraction of the amount of edges and change in the image.

Intensive measurement of the structural information is derived from the correlations between the

amounts of the variation caused in the image. Four correlations namely Pearson correlation

coefficient between Pg and Pl (PRCP), Pearson correlation coefficient between Cg and Cl

(PRCQ), Spearman rank correlation between Pg and Pl (SRCP), Spearman rank correlation

between Cg and Cl (SRCQ) are proposed in this paper.

3.PROPOSED WORK As discussed in the above section, a methodology to find out the profiles of the structural features

and to derive the correlations between the structural features is explained in stages with their

relevant outcome in each stage is given below.

3.1.Local Features Local contrast features give the information of the amount of change in the structure of an image.

Two general local contrast features are Gradient magnitude (G) and Laplacian of Gaussian (L). Discontinuities in the structural details like luminance of an image or the change in the intensities

are important for quality assessment. These can be derived by performing gradient magnitude and

Laplacian of Gaussian operations. We use CSIQ database which is commonly used international

database for image quality assessment. Trolley image is considered from database and denoted by

I. To standardize Convert image from RGB to grayscale. For size of the image, in the following

stages gradient magnitude and Laplacian of Gaussian operators are applied.


195

3.2.Gradient Magnitude Gradient magnitude is the first order derivative, often used to detect the edges in the image.

Expression for gradient magnitude is given as:

G=

The vertical prewitt filter kernel (V) is considered as [-1 -1 -1; 0 0 0; 1 1 1]. The horizontal

prewitt filter kernel (H) is considered as [-1 0 1; -1 0 1; -1 0 1]. Appling these kernels on the

image and substituting in G expression given above, we get gradient magnitude image. The

original image, vertical prewitt filtered image, horizontal prewitt filtered image and the resultant

gradient magnitude image are shown in figure 1.

Figure.1. 1) Original image 2) Vertical Prewitt Filtered Image 3) Horizontal Prewitt Filtered Image 4)

Resultant gradient magnitude image.

3.3.Laplacian of Gaussian

Laplacian of Gaussian is the second order derivative as shown in the equation. Expression of the

Laplacian of Gaussian is given as,

L=I*

Where,

= g(x,y)+ g(x,y)

G and L operations reduce the spatial redundancies in the image. The Laplacian of Gaussian

applied image is shown in figure 2. Some consistencies between neighbouring structures still

remain. So, to remove these we perform joint adaptive normalisation.


196

Figure 2. Laplacian of Gaussian performed image.

3.4.Joint Adaptive Normalisation

Joint adaptive normalization (JAN) is performed to remove the spatial redundancies remained in

the image [1]. This decomposes the channel into different frequencies and orientations.

According to the normalization factor G and L are reduced.

Figure 3. 1) Laplacian of Gaussian 2) Gradient Magnitude 3) Joint Factored Image

3.5.Locally Adaptive Normalization Factor A 3*3 mask which has values which when summated equals to 1 is applied on the image. As the

mask is run over square of joint factored image while finding out the square root of the same,

gives normalization factor. Last step of this procedure is to find out new values of G (i,j) and L

(i,j) as (i,j) and (i, j) by reducing the features by normalisation factor. Variation in Buildings

image before and after joint adaptive normalisation are shown in figure 4.


197

Figure 4. 1) Gradient Magnitude 2) GM after joint adaptive normalisation 3) Laplacian of Gaussian 4) Log

after joint adaptive normalisation

3.6.Quantization The features obtained on applying the techniques of gradient magnitude and Laplacian of

Gaussian are quantised. This is performed to decrease the dynamic range and to bring the features

into an optimum range. We quantized (i,j) into planes as { . Similarly

(i,j)into . In this case we take 17 levels of which we assign 17 different levels of

pixels values. This may be a lossy process but it is done to derive the respective density functions

of gradient magnitude and laplacian of Gaussian features. Resultant Trolley image after

quantization into 17 levels is shown in figure 5.

Figure 5. 1) GM quantised into 17 levels 2) LOG quantised into 17 levels

3.7.Marginal Distributions and Conditional Probability

Dependency measures like Marginal distributions and conditional probability closely relate the

amount of distortion present in the image. This is more comprehensive evaluation of the extracted

features. For the 17 levels of quantized images the marginal distributions and conditional


198

probabilities are derived. To find out the marginal dependencies of and , procedure starts with

deriving the joint empirical functions for all levels.

= , =

Normalised histogram of and is . Marginal distributions of and are given in the

expression below. Marginal probabilities of and are shown in the figure 6.

Pg ( = )=

Pl ( = )=

Figure 6. Marginal distributions of quantised and .

Sometimes the marginal distribution does not show the dependencies between and . The

dependency between them are derived by dependency measure given in equation below.

=

Using the marginal distributions as the weights the conditional probabilities are derived for

and as Cg and Cl. These probability distributions are otherwise called as the independency

distributions. Independency distributions of and are shown in figure 7.

Cg ( = )=Pg ( ) .

Cl ( = )=Pl ( ) .

Figure 7. Independency distributions of and .


199

While finding out the marginal densities of GM and LOG and their corresponding profiles, it is

seen that changes in a distorted image and randomness of distortion is distinguishable through.

Vertical, horizontal and diagonal profiles of and of a not distorted are shown in figure 8. A

normal distortion less image has quite different distributions from that of a distorted version of it.

Similarly their profiles are also plotted. The individual random process of G and L features after

quantization are assumed as the random variables. The below figures fall under the assumption

of binomial distribution of the data. Further there is need to identify the dependency between the

distributions of G and L. Hence, calculating joint probability density function between them is the

solution. In a general case they find to be independent and results are in product of their

individual marginal density functions. For sample profiles over 20 bins are shown in plots in

figure 8.

Figure 8. Horizontal, vertical and diagonal profiles of and .

To present the image in the score accurately, two correlations that can measure scores

that can measure the relation between structure features are used. They are Spearman rank order

correlation coefficient (SRC) and Pearson correlation coefficient (PCC). The dependencies between features of extracted horizontal, vertical and diagonal profiles can

formulate the level of distortion in the image. To find out them, spearman and Pearson

correlations between Pg, Pl and Cg, Cl are calculated. We propose four models Pearson

correlation coefficient between Pg and Pl (PRCP), Pearson correlation coefficient between Cg

and Cl (PRCQ), Spearman rank correlation between Pg and Pl (SRCP), Spearman rank

correlation between Cg and Cl (SRCQ). The scores of four models and their comparison with

Structural similarity value are tabulated in columns below. Highlighted values represent right

ordered values in coincidence with level of distortion.

Scores of AWGN distorted images and their relevant SSIM values are recorded in table 1. For

Blur, Pearson correlations of conditional probabilities give more equivalence. The scores of blur

distorted images and their relevant SSIM values are given in table 2. For AWGN, Spearman

correlations of marginal distributions give more equivalence. The scores of images distorted with

flicker noise and their relevant SSIM values are recorded in table 3. For flicker noise affected

images, Pearson correlations of marginal distributions and conditional probabilities give more

equivalence. The scores of JPEG distorted images and their relevant SSIM value are recorded in

table 4. For JPEG, Pearson correlations of marginal probabilities give more equivalence. The

scores of JPEG2k distorted images and their relevant SSIM values are shown in table 5. For

JPEG2k, Spearman correlations of conditional probabilities and Pearson correlations of marginal

distributions give more equivalence. Overall, Pearson correlation coefficients of marginal

distributions proves to be exemplary.


200

Table 1. Scores of AWGN distorted images and their relevant SSIM value.

Level of Distortion PRCP PRCQ SRCP SRCQ SSIM

1 0.2368 0.9244 0.3333 0.8182 0.9869

2 0.2688 0.9324 0.2001 0.8545 0.9554

3 0.2791 0.9239 0.2256 0.8788 0.8853

4 0.2881 0.9345 0.2121 0.9157 0.7781

5 0.2881 0.9105 0.9761 0.9258 0.6319

Table 2. Scores of Blur distorted images and their relevant SSIM value.


1 0.2299 0.9362 0.4061 0.8182 0.9953

2 0.2149 0.9206 0.4061 0.8303 0.9827

3 0.2045 0.9302 0.4061 0.7455 0.9437

4 0.1775 0.9308 0.4499 0.8909 0.8366

5 0.0941 0.8739 0.6383 0.9031 0.6263

Table 3. Scores of Fnoise distorted images and their relevant SSIM value.


1 0.2577 0.9464 0.2485 0.9031 0.9908

2 0.2689 0.9407 0.2001 0.8667 0.9667

3 0.2781 0.9321 0.2485 0.8909 0.9162

4 0.2808 0.9174 0.2485 0.8788 0.8241

5 0.3162 0.8703 0.3051 0.7939 0.6978

Table 4. Scores of JPEG distorted images and their relevant SSIM value.


1 0.2412 0.9204 0.3333 0.8305 0.9912

2 0.2526 0.8824 0.2485 0.8909 0.9686

3 0.2371 0.9133 0.3708 0.8424 0.9196

4 0.1706 0.8965 0.3212 0.8545 0.7931

5 0.1498 0.8816 0.3697 0.9142 0.6826


201

Table 5. Scores of JPEG2k distorted images and their relevant SSIM value.


1 0.2414 0.9134 0.2918 0.9031 0.989

2 0.2321 0.9064 0.3333 0.8788 0.9637

3 0.1854 0.8377 0.3617 0.9031 0.9008

4 0.1334 0.9458 0.3818 0.8349 0.7839

5 0.0741 0.9203 0.4394 0.8788 0.6097

4. CONCLUSIONS

Existing BIQA models are complex and involve either in exquisite decompositions or model

learning and support vector regression. Few explicit models unlike the proposed method change

the features of the image. Keeping this in concern, an attempt is made to use the correlations

between the statistics of the local contrast features. Since these are independent, data of the image

is not disturbed. In this paper simple procedures to normalise are used to derive joint statistics

with joint adaptive normalisation. Marginal distributions and conditional probabilities and their

dependencies led to highly competitive performance. Avoiding the training and learning of the

features derived, complexity is reduced. Amongst the four models, Pearson correlation coefficient

between Pg and Pl (PRCP) proved to be consistent. However, all the four models have affinity

with structural similarity.While Pearson correlation is a linear correlation, Spearman is a rank

correlation. Hence results are different for different types of distortions in proposed four models

with two correlations because of the variation in structural profiles. This proves that when

variation in the image structure can define the type of distortion present in the image.This can

lead to development of newer models which can determine the type of distortion.

REFERENCES [1] Xue and A. C. Bovik, "Blind image quality assessment by joint statistics of gradient magnitude and

laplacian of gaussian", IEEE Trans on image processing, 2014

[2] Z. Wang and A. C. Bovik, "Mean squared error: Love it or leave it? A new look at Signal Fidelity

Measures", IEEE Trans Signal processing, vol. 26, no. 1, doi.10.1109/MSP.2008.930649

[3] Z. Wang and A. C. Bovik , H. R. Sheikh and E. P. Simoncelli, "Image quality assessment: From error

visibility to structural similarity", IEEE Trans. Image Processing, vol. 13, no. 4, pp. 600-612, 2004

[4] H. R. Sheikh, M. F. Sabir and A. C. Bovik , "A Statistical Evaluation of Recent Full Reference Image

Quality Assessment Algorithms", IEEE Transactions on Image Processing, vol. 15, no. 11 pp. 3440 -

3451

[5] A. K. Moorthy and A. C. Bovik, "A two-step framework for constructing blind image quality

indices",IEEE Signal Process. Lett. vol. 17, no. 5, pp. 513-516, 2010

[6] M. A. Saad, A.C. Bovik and C. Charrier, "A DCT statistics-based blind image quality index", IEEE

Signal Process. Lett. vol. 17, no. 6, pp. 583-586, 2010

[7] X. Marichal, W.Y. Ma and H. Zhang, "Blur determination in the compressed domain using DCT

information", Proc. ICIP, pp. 386-390

[8] A. Mittal, A. K. Moorthy and A. C. Bovik, "No-reference image quality assessment in the spatial

domain", IEEE Trans. Image Process., vol. 21, no. 12, pp. 4695-4708, 2012

[9] J. Canny, “A computational approach to edge detection", IEEE Trans. Pattern Anal. Mach. Intell., vol.

PAMI-8, no. 6, pp. 679-698, 1986

[10] M. A. Saad, A. C. Bovik and C. Charrier, "Blind image quality assessment: A natural scene statistics

approach in the DCT domain", IEEE Trans. Image Process., vol. 21, no. 8, pp. 3339-3352, 2012


202

[11] A. K. Moorthy and A. C. Bovik, "Blind image quality assessment: From natural scene statistics to

perceptual quality", IEEE Trans. Image Process., vol. 20, no. 12, pp. 3350-3364, 2011

[12] D. Marr and E. Hildreth, "Theory of edge detection", Proc. Roy. Soc. London B, Biol. Sci., vol. 207,

no. 1167, pp. 187-217, 1980

[13] G.-H. Chen, C.-L. Yang and S.-L. Xie, "Gradient-based structural similarity for image quality

assessment", Proc. ICIP, pp. 2929-2932

[14] H. R. Sheikh, Z. Wang, L. Cormack and A. C. Bovik, Live Image Quality Assessment Database

Release 2., 2011, [online] Available: online

blind image quality assessment with local ...aircconline.com/ijci/v5n4/5416ijci22.pdfpg scholar,...

Documents